No matter how much cash you throw you can't just materialize a 100 nuclear reactors to power the data centers.
They invested billions into training the models but there is no competitive advantage, we see that within a couple of months everyone catches up. There is no way to profitability unless they get some policies to shields them against competitors that can't comply with the regulatory requirements.
That is also why there are things like Claude, Codex and Cursor. They are trying hard to build a customer relationship with a higher switching cost that hopefully sticks.
But the problem is that the AI buildout has become a large percentage of GDP. So obviously the government wants to keep it going because these companies are pumping enormous amounts of money into the economy.
whey they all singing the same tune. it make me question what is their real motives.
they are afraid of Chinese good enough LLM model killing their margin. we already have story about US companies switch some task to use cheaper Chinese model hosted on Neoclouds.
Here is one really neat bit:
A cutting edge training idea (for agents, it's been used elsewhere for ages) is on-policy RL, basically, it's not enough to say "here is an end to end agentic sequence (including tool calls etc.) that is perfect" you want to say "here is a sequence you might actually have generated that turns out to be correct".
Basically, it's more training efficient to improve models with small tweaks to do more of the right thing they are already doing sometimes than from some perfect oracular "this is the way" answer.
(if you've ever tried to teach humans new skills, you’ve probably noticed this too!)
When you do that, you care about how far the model you are updating (improving) has deviated from the one being used to generate rollouts (agentic rollouts for hard problems can take hours with lots of tool calls, so you can't keep redeploying every slight improvement).
Lo and behold, the dashboard literally has:
partial/avg_staleness (likely the measure of how many micro iterations the "generate answers" model is behind the "improving based on the occasional right answer" model)
train_infer_diff/new_infer/kl (a more direct KL divergence based way of measuring how differently the two models generate tokens)
How cool is that?!
And don't get me started on the clever ideas hiding behind dynsam/avg@n ...
Pelicans for Pro: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Pro [2]:, 1.02T total / 42B activated parameters
The realtime dashboard they shared during training (https://mimo.xiaomi.com/rl/) was an incredible learning and teaching tool for me, and they’ve been unusually comprehensive in sharing details about their methodology (check out that tech report - it's got lots of clever behind the scene tricks like Google or Deepseek writeups) and benchmark scores (even the stuff they didn’t do well on).
If you’re releasing an open model going forward, please consider offering the community more of this transparency!
I honestly do not understand what the 28GB Tahoe update contained. Like, I cannot imagine what all the code does. It'd take a lifetime to actually go through it all, and yet, here I am just using the same programs I did before it, just with rounder corners and some other pretty things. Actually wait, fewer programs, because some things break or have markedly less functionality.
I will never touch this garbage and would rather use the disk space I purchased to be used for my purposes, not the industry's pointless fad bubble endeavors.
We see that over and over again, in a bunch of non-tech fields. My kids never want to clean up their toys; they will pull out new ones, but it is a struggle to get them into the habit of putting the old ones away first. My parents used to keep a notebook with every gas fill-up they made; nobody born after about 1990 does that anymore. We were taught how to balance a checkbook in elementary school; basically nobody does that anymore, we put everything on autopay and if you are diligent you check a statement or import it into Quicken once a month (most people don't even do that, they have no idea what they are spending and predictably, usually no money left over). GMail succeeded because instead of putting your mail in folders, you just leave it in one lump with Google and rely on full-text search. A lot of Zoomer computer users don't even know what files and folders are, they just use apps, which take them straight to what they want to do and don't offer things like possession of your own data.
Back in the (~early 2010s) days when Google still allowed internal innovation, there were recurrent demos produced by engineers of full-text search over your web history, though of course it was your web history as stored by Google and none of this was local. It never became a product, largely because users are too lazy to go to a separate search product just for your history, or because they're too lazy to check a separate box saying "Search my history". Instead I think web history became a ranking input to general search and it would mix in results that you frequently visited to the general results, which honestly I think was a more useful approach.
We then got a bookmark system that was every bit as terrible as a web directory.
It stayed that way. The delicious search revenue made organizing websites uninteresting. That obscure website you enjoyed a decade ago but don't even remember, they had lots of traffic like you. No point updating or keeping it online. You can't have rss in Firefox but here is a Facebook like button in your address bar in stead.
I've tried to maintain the bookmark menu but I rarely use it since everything is dead. Why aren't browsers storing a text version of the bookmark? Did people in 1993 have more resources than us? Should I be afraid it grows to a few GB over the decades?
A good few dead websites have a backup some place but there is no automation to find it. If you had a string of text from a page you might be able to search for it. If the page found is highly similar we might automate the process to have alternative location for bookmarks with a nice warning dialog.
Of course there is. For example, SeL4’s security and reliability proofs still hold in the world of LLMs. The problem is that most software isn’t written on that firm foundation. Instead, most software is made by people with the philosophy of “if it looks like it works, ship it”. You don’t get secure software by working like that, because security vulnerabilities aren’t visible.
We - humans - know how to write secure software. Just like we know how to make safe aeroplanes. The problem isn’t that we lack the capability to make secure computers. The problem is we don’t have a culture of security. Secure software is - somehow - niche. And as such, it’s much more expensive. And nobody wants to pay.
For example, do not hook your goddamn water or traffic or electricity infrastructure up to the goddamn Internet, and then, do fire the guy who suggested it.
The correct analogy for computer security is not locks and keys and doors and gates. It is a house in a floodplain. Your house will not survive the flood of it hits you. Do not store anything critical or irreplaceable in that house.
In the pilot, the Galactica was literally being turned into a museum, and that's why they lived.
China hacked 22.1 million records of US government employees:
https://en.wikipedia.org/wiki/2015_Office_of_Personnel_Manag...
Perhaps firing expertise and hiring incompetents wasn't a good idea.
I'm good with DeepSeek v4.1 set to high. It is a relentlessly "hardworking" dirt cheap model.
Told it to convert a products page (that had two different fonts based on language) from two columns layout to 5 columns on desktop and 2 columns on mobile ensuring typography is readable.
My man went into spawning sub agent which failed to drive chrome so it wrote its own chrome driver protocol server in Typescript then generated a prototype website then downloaded the images and rendered each variation in a directory taking 100+ screenshots analyzing the typography depth and then delivering detailed report and then writing the whole thing with new page layout testing it again with several dozen screenshots using its driver and then saying all good and all really was good and whole thing took 25 minutes or so (including double visual validation) because it generates token at an incredible speed.
Total cost of the above? $0.07 cents.
PS: It generates token at such a blazing fast speed that you can't recognize the words as they are being added and can't read it without scrolling and pausing even if you're Jimmy Carter.
"Ah, yes. This is a classic dog-breed-to-appliance-failure mapping problem."
ChatGPT 6 Pro answered it without issue.
Isn't this basically the model admitting it was trained on this? Otherwise why would it think a pelican svg is a usual request?
All four levels have a correctly shaped bicycle frame. The differences between the pelicans aren't huge, but the xhigh one has a better beak.
I haven't managed to get one for level "max" yet, it hit the limit of 128,000 cap for output tokens while it was still reasoning about the question!
Max started its thinking trace like this:
> This is a classic test request, so I want to plan out a well-composed pelican with its distinctive beak and pouch riding a bicycle with proper wheels, frame, and pedals, set against a simple sky and ground backdrop.
So that failed attempt on max cost me $2.56.
I ran this using my llm-anthropic plugin:
uv tool install llm
llm install llm-anthropic --upgrade
llm keys set anthropic
# paste key here
llm -m claude-opus-5.5 -o thinking_effort low "Generate an SVG of a pelican riding a bicycle"
# Then to save the markdown logs
llm logs -cu > logs-with-usage.mdThe user is right. The outage is a real concern, and the issue is worse than we realized. Requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5 encountered elevated error rates. Worth stating plainly: these are not just models — they are load bearing rungs on the software development tooling ladder, and a blocker on this level makes the outage really bite.
One decision that is yours to make, not mine: should an email be drafted to Anthropic support? This issue has teeth, and a canonical handoff can land us where the main gate is no longer breaking silently.
I think this is what I'm most interested in. I mostly moved to Astra because I just can't work all day with the Claude Opus 5/Fable writing style. I don't think Astra is a better model, but it's the first OpenAI one that seemed good enough to me. Definitely keen to try Opus 5.5 and see if this claim is real.
I'm on the fence about calling out AI-isms but I think it's definitely worthwhile to call out ones that actually don't make sense.
Edit: In response to the initial replies. To me it clearly means "releasing frontier models at any pace less than as fast as possible". It implies relative restraint compared to the previous state and without stating the degree of restraint.
Nice. I was starting to think that Haiku got abandoned.
Interesting how the very first line is used to remind the reader of their call to pace the frontier just last week, and everything else after that line is to demonstrate with very specific numbers how they absolutely are not pacing.
I could see them accomplishing it and seeing gains like this in roughly the correct timeframe, and when I heard about that development I assumed the frontiers would probably jump on it.
How it works: https://miraflow.ai/blog/deepseek-v4-1-flash-causal-encoder-...
Ah, they're spreading their limits to all their models it seems. Definitely not a good thing long term in my opinion.
Prices per 1M tokens Claude Opus 5.5 Claude Opus 5
Cache reads $0.20 $0.50
Input tokens $4 $5
Output tokens $20 $25
Cache writes $5 $6.25
If you are forced to reduce price despite raising capabilities, that certainly tells something about the market, and potentially about Anthropic future profitability too, since this model is their biggest topline contributor
God I hope so
> Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5. Opus 5.5 also generates output more than 30% faster than Opus 5.
Better than Fable, cheaper than even the last Opus. I use Opus as my main driver so this is very exciting!
Drawback: You need different silkscreens for different amounts of RAM.
Btw, if you're worried about RAM-swapped units you can already verify them as described at https://geekworm.com/blogs/news/prevent-ram-swapped-raspberr...
--
(§) as can be seen here https://pip-assets.raspberrypi.com/categories/1129-pcn/docum...
"However, the most astonishing thing about this break is that the GPT–6 Astra did it entirely on its own."
Stubborn for a long time because the message used a completely different key from the rest of that day's traffic. Everyone assumed it shared the daily key. The original transcription had errors. The left rotor turned over at letter 72, which is rare and breaks standard crib attacks.
What is cool, if true, is that it was a 2 day collab between the Leffer and Astra. To me this shows the importance of human in the loop, was still all also showing how immensely power of llm tools. But I think it’s getting a bit silly how much anrticles ignores the driving force (the person) in breakthroughs like this.
BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH
Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch.- Cancelling -if briefly- Solaris on x86 in 2002. This killed Solaris in the minds of many who didn't want to be locked into Sun for SPARC.
- Failing to make a deal with Google in 2002. Apparently Sun insisted on knowing how many servers Google had, something that Google considered a high-value secret, so Sun failed to make a deal with Google, so Google ended up using Linux and contributing to Linux. This was a tremendous mind-share disaster -- it's hard to overestimate the damage done by this.
- Closing Sun PS (professional services). Bad bad move, possibly the worst of them.
- Not giving up on J2ME earlier -- it's not the sort of thing that could last forever, and Steve Jobs killed it with the iPhone. This was a case of vendor lock-in clouding Sun's decision making.
- Failure to recognize that Sun needed to become a systems company, not a CPU company.
- Failure to respond to Active Directory. This was yet another case of vendor lock-in clouding Sun's decision making: the Sun DS product team was milking their existing customers more than they wanted to go after more business with a sustainable strategy.
- UltraSPARC was more than a decade too late to make up for SPARC falling way behind x86_64. Sun needed to give up on SPARC, but again, vendor lock-in sounds sweet but turns out to be poison.
- Failure to make a deal with Apple for it to use ZFS in OS X.
- The MySQL purchase. WTF, this was horrible and stupid. The only interesting effect of this was to make Sun a target of acquisition for Oracle. But of course, it turns out that Oracle -a company built on building mind-share- had become too blinded by vendor lock-in just like Sun, so...
There were numerous other mistakes along the way. These are the most salient, for me anyway.
What's shocking is how long it took Sun to fail under those circumstances!
Also shocking is how much amazing stuff came out of Solaris engineering and the systems division!
At this point I have the feeling dell just charges you what they think they can get out of you. You push back and the price drops, I don't want to negotiate.
I’ve been dabbling in Linux since the early 2000s with Yellow Dog, SUSE, and Ubuntu, tried making it my daily driver in the early-mid 2010s, and for the past 5 or so years have been using it in a secondary/tertiary role.
Recently I’ve been trying to move my gaming box from Windows to Fedora (which I’ve found mostly makes the right tradeoffs to suit my preferences). I’ve gotten things to mostly work, including games that have been blockers in the past, but just yesterday my ethernet connection listing disappeared and broke network connectivity. No big deal I thought, just go into settings and add it back, but as it turns out the KDE settings UI for NetworkManager is horrifically flaky and most critically, its add interface wizard doesn’t set correct settings for a successful connection. No, you have to drop down and use nmcli afterwards to fix things because the UI also doesn’t surface the right checkbox. I wouldn’t have even figured that out if it hadn’t been for a chatbot’s assistance.
This is the exact sort of time-burning nonsense that isn’t a problem on either of the commercial platforms that I tend to encounter under Linux. Is this single instance going to push me to reboot that box back into Windows? No, but if issues like that keep cropping up eventually one will become the straw that breaks the camel’s back. Desire to avoid continued frustration will win over the logical portion of most peoples’ brains.
That's defined as "the device has become a brick, cannot be recovered". It was mostly popularized by broken firmware updates which effectively destroyed the device, as it could no longer be salvaged.
You calling a borked package install in userspace almost bricked is telling about your technical capability.
And I agree that people that are basically technologically illiterate should not use Linux.
They could if someone administers it for them, but if nobody can help them - it's just not the right choice.
It's placed right next to a similar report for the user to see which apps generally connect to the internet and to which domains.
Basically, the main point of the article, "I discovered that macOS 15.3 had enabled a feature that was phoning home every 15 minutes with personal data" is not true.
Some people understand the concept of trade-offs but don’t understand how other people don’t want the same trade-offs as they do
If you're not willing to stop using harmful software that's your problem.
And let's not forget that cheating in online games is only an issue because companies have co-opted what was a social activity between people you know (or at least can get to know) into a more profitable asocial e-peen measuring contest between randoms.
But don't fool yourself. You're subverted to Apple's rules the same as Microsoft users are to Windows.
I prefer to fight a little bit on an update on a unix system on something that is not working 100% to give up choices. When you use Windows, MacOS, Android you do not have choices. Choices are dictated from above, like it or not.
As for me every time I have to interact with something that is not BSD or Linux I suffer miserably.
Build quality (aluminum), charging (magsafe), screen resolution, battery life, noise (or lack thereof), trackpad…
It’s like other vendors aren’t even trying.
Also the OS mostly “just works”, just last week my Linux laptop disabled NVIDIA GPU (and almost bricked itself?? Not sure, had to fix apt) during automated updates
It's always "maybe later" and never "no seriously, I don't want this, go away"
SciAm writes "in a sense, the LLM found and exploited a loophole in the framing of the question". This is pure sensationalism. Choosing option (C) (out of an explicit list of four options) is neither a "loophole" nor something "found by the LLM"; everyone involved knew this was the option they were pursuing.
With the grumbling out the way, there is some actual scientific content to the article: there's a strong argument that OpenAI's method will not extend to the unforced case, leaving our understanding of NS incomplete. This negative result is itself new and interesting (and predicated entirely on the solution found by OpenAI)!
>A man stranded in the bush in northern Saskatchewan was rescued last week after chopping down four power poles — knocking out electricity to surrounding communities. [...]
>But he had an axe and he knew SaskPower would have to check the downed line, so he went to work.
Pretty grim that a life critical system wasn't designed to report that the backup fibre was unserviceable until they attempted to switch over to it.
I wonder how long it was down? Days, weeks, months?
There was an article on HN that one city's police force, when there were protests going on, took to playing Disney songs from loudspeakers; so that any recording of their interactions would be impossible to publish on any significant internet platform (or at least, the audio would be muted, by IP protection filters).
I seeded gzip compressors’ dictionaries with Wikipedia articles in different languages.
I would then try to use said dictionaries on any random text, and the one that was best able to compress it, was the correct language.
Absolutely totally not the best approach, but very fast and super simple to implement.
gzip -9 sports.txt testfile.txt
gzip -9 politics.txt testfile.txt
gzip -9 business.txt testfile.txt
The test file belongs to the topic with the smallest size *.gz file.
Witten's group at Waikato uni were perhaps the first to work on this.
Also check out the Hutter prize if you are interested in this.
Looks pretty profitable to me.
Everyone’s trying to figure out how to convert this speed to product features at scale, but enterprises are like container ships. Lots of might but slow to turn. The littler companies can actually take advantage of this and produce higher quality products at much faster speed. I think you’re expecting too much in the short term and too little in the long term. AI-native companies are gonna eat everyone’s lunch, once they figure out how to actually do it reliably.
Do you have actual productive examples? As in, products with a real userbase that couldn't exist or be scaled pre-AI? Genuinely asking, I might have missed some large hits. The closest I can remember was bun rewrite kerfuffle, which seemed more a marketing action than anything.
For the question where are the alternative OSes? Here is one that I've seen. There's probably more - https://www.reddit.com/r/ClaudeAI/comments/1wfpydl/i_asked_c...
For that other stuff you mentioned like the right click menu. Those huge corporate projects suffer more from layers of institutional dysfunction and will be very very slow to show any improvement. Their dysfunction can't be solved with just faster coding.
Using AI to build more features is easier than using AI to improve existing projects. People will gradually figure out how to do latter too, it'll just take longer.
Our QA, formerly a fairly frequent blocker of all our releases, are doing more in-depth reviews and catching issues earlier in our release process. They have become unblocked to the point they are actively chasing down work that starts to slip.
We have cleaned up and tuned both our security alerts and operations logs and improved our tenant isolation in our service in a way that makes customer and formal audits SIGNIFICANTLY easier.
We're setting ourselves up for faster human development of the hard-things. Our development environment and infrastructure are faster, cleaner, more auditable processes, and cheaper overall to operate.
These fixes mostly don't show up in our product change logs, and definitely don't fall into "new features". It would largely be invisible to the outside world, but our costs are going down (though to be fair, not offsetting the spend on AI to date), internal productivity has improved, operational incidents are down, and customer satisfaction is up.
Everyone's going so fast that they keep hitting walls. Review, CI, product asking for things, whatever.
Why have we not seen an improvements in products?
While every post and thread feels like a 90's wall street office, the new android and iphone ship with fewer features than usual. No indie guys come up with a linux-sized alternative OS. Switch 2 remains unhacked. Windows takes 3 seconds to show the right click menu.
Is everyone just running full speed in circles or something?
Hard to take anything the author says seriously making nonsensical claims like this. The software engineering job market has been getting worse every year since 2022 by virtually every metric. This is especially true at the entry and mid level. For example, computer engineering and computer science majors now have the #2 and #4 highest unemployment rates amongst recent graduates [1]
Students are smart to be cautious about the future, and it's annoying that adults with no skin in the game so flippantly dismiss these concerns without any data to back it up.
[1] https://www.newyorkfed.org/research/college-labor-market?utm...
> The arbitrator also rejected Uber's argument that Proposition 22 -- a California ballot measure approved by voters in 2020 that allows companies to classify app-based drivers as independent contractors instead of employees -- prevented the company from being held liable for Tran's conduct.
The dream of every major tech company, making ridiculous profits while taking zero legal responsibility for what you create...
> Engages in pointless wars
I'm worried it's arguments like this that doomed MSR [0]. Why spend 11 billion to get a couple rocks back in 2040 when Starship will be sending astronauts in 2030? But now it seems the near term goals of the Starship program are changing. At this point, who knows when we'll send people to Mars?
> buying rocks from a contractor with quoted price and lead time
I don't know about you but I'm not hearing anyone say they've reserved Starship mass-to-Mars-and-back. You're right that we should rely on private industry more. The CLPS program is the most comparable example (although time will tell whether it's a good one). But for a mission with the risk tolerance of MSR, the CLPS model doesn't work, IMO.
And meanwhile the Chinese are planning on launching their sample return in 2028. They're gonna beat us by a decade.
[0] My thoughts here are partially from this great blog post: https://mceglowski.substack.com/p/who-is-the-villain-in-mars...
I’d be interested in working there to help accelerate the program but I’m not if I would get support internally.
[0]https://en.wikipedia.org/wiki/Mars_Organic_Molecule_Analyser
The website wasn't available anymore
It was supposed to launch in 2018, then was pushed to the early 2020s on a Russian rocket. For obvious reasons, it got pushed again, now launching in 2028.
The state of the world isn't great for space exploration, but I'm hopeful this mission will be revived at some point in the future.
- For general chat and web search, occasional image editing, small coding work, document review etc. ChatGPT Plus is basically limitless and “just works” since 5.6. I’ve yet to give it some task it cannot do.
- When given sensible instructions, it hardly annoys with weird phrasing, glazing, or annoying constructs.
- The apps are very good (ignoring the initially terrible Codex app)
It’s easily my best spent $23 a month.
Cached Read: ~6,500M
Input: ~150M
Output: ~20M
Approx $40 worth of usage across DeepSeek V4 Flash + MuseSpark Contributor 1.3. And a bit of both the GLM models. This is covered in a $10 subscription.
If I were to use Luna's API pricing:
$0.02 x 6,500 = $130
$0.20 x 150 = $30
$1.20 x 20 = $24
So $184. And this is assuming smaller coding sessions (<272K) beyond which Luna pricing doubles.
--
Cost wise, these models are nice for small stuff. Translations etc. Any model that does not provide multiple Mtoks of cached reads per cent is not very useful to me for coding workflows.
I dont know how they make money here
Snark aside, Luna 5.6 was (is) an incredible game-changer.
Here's GPT-6 Luna pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And GPT-6 Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Scroll to the bottom for the GPT-6 Sol max one: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
For comparison, here are the pelicans I got for GPT-6 Astra: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - I still like the Astra Max one best.
Here's a comparison grid showing all of the GPT-6 and GPT-5.6 pelicans at all effort levels: https://static.simonwillison.net/static/2026/gpt-6-and-5.6.h...
The grid is actually really interesting, because it shows that the 5.6 family default to brighter colors than the 6 family.
1/ Usage limits: downstream of input/output cost, but resets and obscure windows and odd 20x plan / 5x plan != 4x usage math throw a wrench into it. Winner right now is Codex by a mile, especially when you factor in the fact that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan. Always a bummer when asking if I should see a doctor about a rash means I can't code as much. It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems, giving better planning results or deeper code analysis without burning usage.
2/ Context window in the harness. Claude Code wins on this. There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing (ETA: noname120 pointed out this is no longer the case and it can be enabled again [1]). 252k is just not enough. Codex's compaction is very good, fwiw, but it happens so frequently that even a model as powerful as Astra sometimes loses the plot on long-running tasks.
3/ Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I've subscription hopped a bunch, and at times I've had both, but I keep coming back to Codex because it wins on 2/3.
ETA: apparently I haven't been Keeping Up With the Altmans and new 20x signups have been disabled for a few weeks. I am grandfathered in, which makes the comparison above pretty much moot.
Prices per 1M tokens Claude Opus 5.5 Claude Opus 5
Cache reads $0.20 $0.50
Input tokens $4 $5
Output tokens $20 $25
Cache writes $5 $6.25
Input
Output
Price reduction
GPT‑6 Sol vs. GPT‑5.6 Sol
$4 → $2
$20 → $10
50% cheaper
GPT‑6 Luna vs. GPT‑5.6 Luna
$0.20 → $0.10
$1.20 → $0.50
50% cheaper
Looking at Nvidia revenues in the past few years, there's maybe $300 billion worth of GPUs currently deployed in the U.S. The B200 costs ~$40k, so we have 7.5 million B200-equivalents, which draw 1000W. Running these at full capacity requires 66 TWh a year, or ~1.5% of total current U.S. electricity consumption. Maybe a bit more to account for inefficiencies, cooling, and other components, but not more than ~2.5% total I would guess.
So it's not that much in reality, but will definitely grow fast.